Differences

This shows you the differences between two versions of the page.

Link to this comparison view

a_novel_nonlinear_approach_to_speech_perturbation_measure_for_pathological_voice_classification [2014/04/02 22:34] (current)
bziolko created
Line 1: Line 1:
 +Khalid Daoudi, Safaa Mrad (France)
  
 +**A novel nonlinear approach to speech perturbation measure for pathological voice classification**
 +
 +INRIA Bordeaux
 +
 +Standard speech perturbation measures such as Jitter, Shimmer and HNR are generally based on a pitch marks detection algorithm which assumes the existence of a periodic pitch pattern and/or relies on the popular linear source-filter speech model. While these assumptions can be valid for normal speech, they are generally not valid for pathological speech. The latter can indeed present strong aperiodicity, nonlinearity and turbulence/noise. This paper proposes new definitions of Jitter, Shimmer and HNR which do not make any periodicity nor linearity assumptions. Our approach is based on a novel algorithm for Glottal Closure Instants (GCI) detection that we have recently developed and which outperforms state-of-the-art methods, particularly in the presence of noise. The principle behind this nonlinear and multiscale algorithm is the detection of critical transitions in complex signals (such as speech). As such, the algorithm processes speech as a nonlinear dynamical system without prior hypothesis. We first use this algorithm to define “Critical Transitions Marks (CTM)” and show that they coincide with pitch marks (up to a shift constant) for normal speech. However, for pathological speech, they are completely different from the pitch marks provided by standard algorithms and correspond to real regime transitions in the speech signal. We then use these CTM as the core of new definitions of Jitter, Shimmer and HNR. We carry out experiments on the full KayElemetrics database of sustained vowels. We first show that, for normal speech, our new perturbation measures coincide with those of the MDVP and Praat softwares. We then compare the normal-vs-pathological classification performances. The results show that every new measure significantly outperforms its MDVP/Praat counterpart. Our algorithms will be freely available online
Copyright © XXII PVC Organizing Committee 2013. All Rights Reserved.